search

microsoft / Pengi

An Audio Language model for Audio Tasks
https://arxiv.org/abs/2305.11834
MIT License
284 stars 15 forks source link

the files number of TUT 2017 dataset #12

Closed Zth9730 closed 8 months ago

Zth9730 commented 9 months ago

The Table 2 in the paper says that TUT 2017 contains 6.3k files and includes training, testing, and validation sets. But when I download TUT 2017, there are only 312*15=4680 files and there is only development set. May I ask why there are 6.3k number of files here and how it is divided into training testing validation set? image

soham97 commented 9 months ago

Hi @Zth9730, TUT 2017 has a total of 6.3k files. It's divided into development and evaluation and can be obtained from the below links:

Table 2 in the paper shows full dataset statistics. For Pengi's zero-shot evaluation, we use only the evaluation set to ensure our numbers are comparable with other zero-shot and supervised benchmarks. I hope this helps!

Zth9730 commented 8 months ago

Thanks a lot !!!